“Q 分布”神经网络的输出：

out = layers.fully_connected(out, num_outputs=num_actions * nb_atoms, activation_fn=None)

out = tf.reshape(out, shape=[-1, num_actions, nb_atoms])
out = tf.nn.softmax(out, dim=-1, name='softmax')

选取 Q 值最大的为最佳动作：

def p_to_q(p_values, dist_params):
    z, _ = build_z(**dist_params)
    print(z, p_values)
    return tf.tensordot(p_values, z, [[-1], [-1]])

def pick_action(p_values, dist_params):
    q_values = p_to_q(p_values, dist_params)
    deterministic_actions = tf.argmax(q_values, axis=1)
    return deterministic_actions